Turn a Great Chat into a Reusable Skill

You will produce a Skill Extraction Worksheet, a draft skill package, and a three-case evaluation record.

People with a successful AI conversation who want reuse without copying an entire transcript into every new task.

• A conversation and final artifact you may reuse
• A clear reason the result was successful
• A redacted copy of the conversation
• A target agent environment
• Two synthetic examples and one out-of-scope case

1. 1. Redact the source

   Remove secrets, private names, customer data, paid-source passages, and unrelated personal context. Search the transcript and filenames, not only the final artifact.

2. 2. Mark the turning points

   Highlight instructions, examples, questions, corrections, and tool use that materially changed the output. Ignore conversational politeness and repeated context.

3. 3. Separate method from luck

   For each turning point ask whether a fresh agent needs it, whether it generalizes, and what evidence proves it mattered. Convert durable moves into ordered instructions.

4. 4. Write the contract

   Define trigger, non-trigger, inputs, output, permissions, stop conditions, and definition of done. A trigger should be recognizable from the user's request.

5. 5. Package progressive detail

   Keep the main skill concise and navigable. Put long references, templates, and examples into clearly named supporting files. Do not include scripts without review.

6. 6. Test in a clean task

   Run a normal case, an incomplete case, and an out-of-scope case. Score the artifact against the same observable criteria as the original success.

7. 7. Decide adoption scope

   Record provenance, license, version, files, permissions, target agents, and known failures. Installation remains a separate approved action.

Synthetic example

Scenario

A conversation produced a strong research brief after several human corrections.

Input

The corrections required an explicit decision, primary-source hierarchy, claim statuses, contradiction handling, and a reversal condition.

Result

A Research Brief skill that triggers on consequential research requests, refuses to draft before a claim ledger, produces a brief plus ledger, and correctly declines a request for unsourced promotional copy.

• Copying the transcript instead of extracting a method
• Leaving private context inside examples or metadata
• Writing a trigger so broad the skill runs everywhere
• Testing on the original input only
• Treating packaging as permission to install or execute scripts

• □The source has been redacted
• □Each instruction traces to a demonstrated need
• □Trigger and non-trigger are distinguishable
• □A clean agent succeeds without the original chat
• □Incomplete and out-of-scope cases fail safely
• □Provenance, license, version, permissions, and target are recorded

skill-extraction-worksheet.md

# Skill Extraction Worksheet

## Source and rights
- Conversation owner:
- Reuse allowed:
- Redaction completed:

## Demonstrated success
- Artifact:
- Observable quality criteria:
- Important human corrections:

## Skill contract
- Trigger:
- Do not trigger when:
- Required inputs:
- Output artifact:
- Tools and permissions:
- Stop conditions:
- Definition of done:

## Durable method
1.
2.
3.

## Package plan
- Main instructions:
- References:
- Templates/examples:
- Scripts requiring review:

## Evaluation
- [ ] Normal synthetic case
- [ ] Incomplete-input case
- [ ] Out-of-scope case

## Adoption record
- Version:
- License/provenance:
- Target agent:
- Known limitation:
- Installation approval: Not granted by this worksheet

1. 0:00 Show the successful artifact, not the chat interface
2. 0:35 Redact the source and mark a turning point
3. 1:25 Separate durable method from one-off wording
4. 2:15 Write trigger, output, and stop condition
5. 3:05 Show the package file plan
6. 3:40 Run the normal and out-of-scope tests
7. 4:35 Record the installation boundary